Incoherent, broadband, large-aperture metasurface edge encoders

The high computational demand of deep neural networks for computer vision must be alleviated by hardware-based strategies to facilitate applications in resource constrained systems. A potential solutions is to offload computations onto a front-end analog optical preprocessor, which could perform low-level feature encoding operations instantaneously as images are captured. To that end, this study showcases incoherent, broadband, low-noise optical edge encoding for thermal imaging of real-world scenes, which is achieved using a hybrid system of a 24-mm, inversely-designed metasurface and a refractive lens. Using an inverse design approach, the metasurface is optimized for Laplacian-based edge detection across the 7.5 – 13.5 µm LWIR imaging band. This work could be expanded to enable optically-encoded feature maps for accelerating convolutional neural networks for image segmentation and classification.